Lecture 8: The Ising Model and Mean-Field Theory

8.1 Recommended textbook chapters for this section

• Chandler - 5.1, 5.4

8.2 Topics in this lecture

• The Ising model

• Mean-field theory

8.3 The Ising model

In the last lecture, we finished our derivation of the ideal gas partition function and used its
form to derive well-known thermodynamic expressions (such as the ideal gas of equation of state).
We also derived a Langmuir adsorption isotherm following a similar procedure. In both of these
cases, we neglected interactions between molecules. The approximation that molecules do not
interact is severely limiting and makes it impossible to analyze most condensed phase systems,
such as liquids, or examine phase transitions, which fundamentally depend on the strength of
intermolecular interactions. In general, studying particles with interactions is very challenging,
leading to very few problems for which exact analytical solutions exist. This difficulty emerges
because interactions between particles lead to correlations between particle configurations; that is,
the partition function can no longer be factorized into individual single-particle partition functions
because the energy of one particle depends on the configuration of the others. In this lecture, we
will focus on a simple type of lattice model - resembling lattice models we have treated previously -
to illustrate how interactions affect many-body systems, and to introduce a common approximation
scheme for studying these systems.
The system we will consider consists of a set of N magnetic spins placed on a lattice, to represent
a magnetic system. The lattice can be of various dimensions and geometries (e.g. 2D square or
hexagonal). Each spin can either be up or down; we denote the value of spin i as si = 1 for spin
up and si = −1 for spin down. The lattice may also be in the presence of magnetic field, H, that
tends to align spins with the field. The energy of a single configuration of this system is:
N
X
Eν = − µHsi + (energy due to interactions between spins) (8.1)
i

µ is the magnetic moment (not to be confused with the chemical potential). Each microstate of
the system refers to a particular set of spins - that is a set of values of s1 = ±1, s2 = ±1 . . . , which

1
University of Wisconsin-Madison Lecture 8
CBE 710, Fall 2019 - Prof. R. C. Van Lehn October 1, 2019

we will denote as {si }. In Problem Sets 1 and 2, we already looked at the behavior of a system of
non-interacting spins in which the latter term in Eq. 8.1 is ignored. Here, we will consider what
happens when we incorporate these interactions.

A simple model for the interaction energy is to assume that only nearest-neighbor spins interact,
and that this interaction can be described by a coupling constant J. A pair of spins contributes
energy −J to the system if they are aligned, and a pair of spins contributes energy J to the system
if they are oppositely aligned. The total energy of the system can then be written as:
N N 0
X J XX
Eν = − µHsi − si sj (8.2)
2
i i j

Here, i runs over all spins, while 0j is only over the nearest-neighbor spins of spin i, indicated
P P
by the 0 symbol. The negative sign in front of the J is because the product of si sj is positive for
aligned spins, but the energy between aligned spins is negative. Because the outer sum runs over
all spins while the inner sum runs over all neighbors, the interaction between spin i and j will
be counted twice - once when considering the neighbors of spin i, and once when considering the
neighbors of spin j. We divide by 2 to remove this overcounting of interactions (this is often a
confusing point that is neglected by some authors in their treatment of this model).
This expression for the energy of a set of spins on a lattice is referred to as the Ising model.
The behavior of the Ising model is largely determined by the value of the coupling constant, J. If
J = 0, we regain the expression for non-interacting spins studied in the problem sets. If J > 0, it is
energetically favorable (i.e., the energy is minimized) if all spins align - that is, for all neighboring
spins to have the same value of s. Conversely, if J < 0, it is energetically favorable for spins to
have alternating alignments. For this system, the total magnetization is:
N
X
Mν = µsi (8.3)
i

The system has a spontaneous magnetization if hM i = 6 0 in the absence of a field (H = 0).

If J > 0, then in at equilibrium the lowest-energy configuration would align all of the spins and
there would be a finite magnetization in the absence of a field. We call this system ferromagnetic.
Conversely, if J < 0 and spins align in alternating orientations, the spins would be ordered but
there would be no net magnetization. We call this system antiferromagnetic. Finally, if J = 0
and spins are randomly oriented in the absence of a field, we call the system paramagnetic.

2
University of Wisconsin-Madison Lecture 8
CBE 710, Fall 2019 - Prof. R. C. Van Lehn October 1, 2019

Note that aligning spins reduces the entropy of the system. For example, consider a case
when J  0 and all spins are in the same direction. The system can only have two possible
microstates - all spins are up, or all spins are down. In contrast, a disordered alignment of spins
would have many more possible configurations. There is thus a trade-off between entropy and
energy in systems for which J 6= 0. Consider a ferromagnetic system (J > 0) at constant N V T ,
with H = 0 and finite T . At equilibrium, the Helmholtz free energy of this system must be at a
minimum. Writing F = E − T S, the system can switch between a low energy state, in which spins
align due to the positive value of J, or a high entropy state, where spins are randomly aligned.
Therefore, there should be a temperature at which there is an order-disorder transition and the
magnetization changes from zero (at high temperature) to a non-zero value (at low temperature)
due to the energetic interactions between spins. We call the temperature at which this order-
disorder transition occurs the Curie temperature, TC , for this magnetic system, and will now study
its origin. However, it is important to note that while we are strictly deriving this for a magnetic
system, it will turn out that the same transition is analogous to phase transitions in other systems,
such as the liquid-gas phase transition. For example, instead of treating a series of spins that are up
or down, we could apply the same formalism to a series of particles that are of one type or another
(e.g., to represent the mixing of two types of liquid-phase molecules; this is called the lattice gas
model) which would lead to similar behavior.

To determine an expression for the Curie temperature, we need to relate the ensemble-average
magnetization of the system to the temperature. As our system can be described by the canonical
ensemble (if we include the magnetic work term as part of the energy function, as done above), our
first step will be write an expression for the canonical partition function. Using the expression for
the energy, the canonical partition function for the Ising model can be written:

3
University of Wisconsin-Madison Lecture 8
CBE 710, Fall 2019 - Prof. R. C. Van Lehn October 1, 2019

X
Z= exp (−βEν ) (8.4)
ν
 
N N X 0
X X J X
= exp β µHsi + β si sj  (8.5)
ν
2
i i j
 
N 0
N X
X X X X J X
= ··· exp β µHsi + β si sj  (8.6)
2
s1 =±1 s2 =±1 sN =±1 i i j

The last expression explicitly sums over all possible configurations of the {si }, where for each
set of spins the energy is defined by summing over that set. This partition function is much more
complex than the versions we’ve seen to date since we cannot factorize it in any obvious way
due to interactions between spins. It should thus be apparent why systems with interactions are
challenging to solve analytically.

8.4 Mean-field theory

Analytically solving the Ising model partition function is possible in a one-dimensional system
(Problem Set 3!), possible in a two-dimensional system with significant difficulty (first solved by a
Nobel laureate), and impossible in higher dimensions. Thus, we will instead simplify the partition
function and introduce an approximation that is commonly used in the study of phase behavior
- the mean-field approximation, or the approximation used in mean-field theory. The idea
of mean-field theory is to examine a particular particle (for the Ising model, a single spin) and
assume that the surrounding medium (in this case the other spins) exerts an average field which
interacts with the tagged spin of interest. That is, instead of examining interactions with the
specific configuration of neighbors local to that spin, the interactions are assumed to be equivalent
to that of the average configuration of the medium. The approach thus neglects fluctuations of
the medium that are long-ranged and reduces the many-body problem to an essentially one-body
problem focused on the particle of interest. We will illustrate the essence of mean-field theory by
using it to analytically solve the Ising model.

Consider again the energy of a single configuration of the Ising model:

4
University of Wisconsin-Madison Lecture 8
CBE 710, Fall 2019 - Prof. R. C. Van Lehn October 1, 2019

N N 0
X J XX
Eν = − µHsi − si sj (8.7)
2
i i j
 
N 0
X J X
=− si µH + sj  (8.8)
2
i j

This expression is for the energy of the entire lattice. If, instead, we consider the energy of a
single spin i, we could write:

0
X
i = −µHsi − Jsi sj (8.9)
j
 
0
J X
= −µsi H + sj  (8.10)
µ
j

Note that the factor of 1/2 is removed from in front of the sum, since we are considering the
total energy of the single spin in the system, including all interactions with neighbors (again, often
a point of confusion). We would need to divide the neighbor interaction energy by 2 if summing
over the entire lattice. The form of this energy suggests that we can think of the term Jµ 0j sj as
P

a perturbation to the actual magnetic field, H, acting on the spin; we therefore define Hieff as the
effective magnetic field that does work on spin i including this perturbation:

0
JX
Hieff ≡ H + sj (8.11)
µ
j

The single-spin energy is then written as:

i = −µHieff si (8.12)
We can now see that the interaction energy with neighboring spins contributes to the effective
field that acts on spin i. We can rewrite the expression for the effective field in terms of the average
spin of the nearest-neighbors to remove the sum:

J
Hieff = H + nhsi0i (8.13)
µ
Here, we define n as the number of nearest-neighbors (i.e. 2 in 1D, 4 in a 2D square lattice,
etc.) and hsi0i is the average value of the spin of the nearest-neighbors of si , hence the 0 .
We now make the mean field assumption, and assume that hsi0i = hsi - that is, the average value
of the local spins (nearest-neighbors) of spin i is equal to the average value of spins in the entire
system. The effective field then becomes:

J
Hieff = H + nhsi ≡ H eff (8.14)
µ
i = −µH eff si (8.15)

5
University of Wisconsin-Madison Lecture 8
CBE 710, Fall 2019 - Prof. R. C. Van Lehn October 1, 2019

These expressions are the essence of mean-field theory - the specific accounting of interactions
with nearest-neighbor spins has been replaced with an average (mean) interaction, H eff , so that
spins are now independent. The mean field is identical for each spin and does not depend on si .
Correlations between spins are completely captured in the ensemble-average spin term, which is
a property of the state of the lattice and not a property of a specific spin. We can then use this
expression for the energy of a single spin to solve for the magnetization. Indeed, we have already
done just that on Problem Sets 1 and 2, where we solved for the magnetization of a paramagnet in
a field H. We can use the exact same formalism but replace H with the effective field H eff :

Z = zN (8.16)
X
z= e−βi (8.17)
i
eff eff
= eβµH + e−βµH (8.18)
eff
= 2 cosh(βµH ) (8.19)
h iN
Z = 2 cosh(βµH eff ) (8.20)

We have simply replaced the field H with the effective field H eff in these steps. The rest of the
derivation proceeds the same as in the problem sets, allowing us to skip to:

hM i = N µ tanh(βµH eff ) (8.21)

= N µ tanh(βµH + βJnhsi) (8.22)

This is the mean-field approximation for the total magnetization of the lattice, M , and
allows us to determine the magnetization for a set of interacting spins without knowing exact
configurations of spins on the lattice. We can simplify this expression by replacing the ensemble-
average magnetization with the ensemble-average spin via hsi = hM i/N µ to finally get:

hsi = tanh(βµH + βJnhsi) (8.23)

This equation has hsi on both the left hand side and in the tanh function. We refer to this type
of expression as a self-consistent mean field theory, because the mean field which influences the
average value of the magnetization then itself depends on the average value of the magnetization.
For this type of expression, we can solve for the solution graphically by finding the intersection
between hsi and tanh(βµH + βJnhsi) if both are plotted as a function of hsi. In other techniques,
such as the self-consistent field theory of polymers, similar equations are solved iteratively by
converging onto a result.

6
University of Wisconsin-Madison Lecture 8
CBE 710, Fall 2019 - Prof. R. C. Van Lehn October 1, 2019

Performing this graphical solution method shows the following behavior in the zero-field (H = 0)
limit. For small values of βJn, the only solution is the trivial solution hsi = tanh(βJnhsi) = 0. Since
βJn is temperature dependent through the β parameter, we recognize that this solution, in which
there is no spontaneous magnetization, is the disordered state. Since β = 1/kB T this corresponds to
a high-temperature state, consistent with our understanding of the entropically favored disordered
state being preferred at high temperatures. For large values of βJn, there are three solutions for
hsi; the trivial solution, and two symmetric solutions for finite values of hsi. The two possible
finite values correspond to the ordered, low-temperature alignment of spins characteristic of a
ferromagnetic, with two solutions possible due to the symmetry of the system. Calculating the
free energies of these solutions would show that the finite solutions are lower free energy than the
trivial solution, confirming that at low temperatures the system aligns spins. Finally, the crossover
between the low-temperature and high-temperature state occurs at a temperature:

nJ
TC = (8.24)
kB
Thus, the mean field theory predicts an order-disorder transition. However, it should be noted
that this solution is not identical to the exact solutions - for example, on a 2D square lattice, the
Onsager analytical solution predicts TC ≈ 2.3J/kB which is distinct from the value of TC in the
mean field model. This error decreases as we go to higher dimensions, but reflects the approx-
imations within mean-field theory. More advanced techniques are possible, including “tagging”
multiple spins for simultaneous analysis, but that is beyond the scope of this class.